Popper held that falsifiability is what demarcates science from pseudoscience. Pseudoscientific theories are compatible with all possible observation. By contrast, scientific theories “stick their necks out”, where certain observations falsify the theory.
There have been various criticisms of Popper’s falsification criterion:
- There is no distinction between observation and theoretical statements, because observation is theory laden. Popper realized this. He held that it was up to the scientific community to decide on what counted as observation statements.
- If a statement S is falsifiable, we can conjoin any pseudoscientific statement P, so that S&P is falsifiable. Presumably if P is not scientific, then S&P should not be either. This is the tacking problem.
- Popper’s falsification has the odd consequence that a statement can be scientific while the negation is not. While “all swans are white” is falsifiable by the observation of a single non-white swan, the negation of that statement “there exists a swan that is non-white” is not falsifiable by finite observation.
- As the Duhem-Quine thesis points out, all theories come with auxiliary theories. In order to deductively falsify a theory, you need to verify the auxiliary hypotheses.
- Probability statements are not, strictly speaking, falsifiable. A fair coin is not necessarily falsified by 100 heads in a row. Popper’s solution was to say that scientists were to specify, in their own fields, how improbable a result must be for a theory to be falsified.
Popper’s falsification criterion makes theories vulnerable, but it has the criticisms mentioned. We want theories to be vulnerable to observation. Theories could not do explanatory work if they were consistent with every possible observation. Elliott Sober proposes an alternative to Popperian falsification that maintains the scientific virtue of vulnerability; and it follows from the Likelihood Principle.
Likelihood Principle: O strongly favors H1 over H2 if and only if P(O|H1) >> P(O|H2)
The likelihood P(O|H) is not to be confused with the probability P(H|O). Suppose O is “there is a noise in the attic”, and H is “there are gremlins bowling in the attic.” While P(O|H) is high, P(H|O) is low. It would be strange to say that we should believe gremlins are in the attic because the likelihood is high. The likelihood, by itself, does not tell you which hypothesis is more plausible. It simply tells you that the observation favors one hypothesis over the other. The plausibility of the hypothesis will be a function of the likelihood and the antecedent plausibility.
So how can the Likelihood Principle make theories vulnerable? If P(O|H1) > P(O|H2), then P(not-O|H1) < P(not-O|H2). This is derived from that fact that P(O|H1) + P(not-O|H1) = 1. In other words, if O favors H1 over H2, then not-O favors H2 over H1, hence the vulnerability. For Sober, testing is contrastive: testing a hypothesis means testing it against another hypothesis.
Sober, Elliott. Philosophy of Biology.
Sober, Elliott. Testability.